The frontier question in robotics right now is not how a machine walks — it is how many hours of somebody else walking it the model has seen. A vision-language-action (VLA) model, the policy stack now bolted into most serious humanoids, learns the way a large language model does: from scale. And the largest teleoperation dataset anyone has admitted to — roughly 80,000 hours of remote-operated mobile-manipulation logs — was just put on the table in Tokyo as a shared, precompetitive commons.

That offer came from Tetsuya Ogata, who directs Waseda University’s Institute for AI and Robotics and chairs the nonprofit AI Robot Association (AIRoA). Waseda is where the humanoid was born: WABOT-1, 1973, the first full-scale bipedal robot that walked, grasped, and spoke a little. Ogata’s argument, as reported by Tim Hornyak in IEEE Spectrum, is blunt — AI is a game of scale, Japan can still secure baseline scale in data, and its only route back is to stop hoarding that scale inside one company and pool it instead.

Why the humility? Because the floor of the Humanoids Summit told the story without a slide. Of about 40 robots on display, Chinese systems outnumbered Japanese roughly three to one; some Japanese firms demoed on Chinese hardware, and one engineer called it “sad.” The comparison everyone reached for was Asimo versus the Unitree G1. Honda’s Asimo was a breathtaking technology demonstrator, never commercialised, retired in 2022 — the year ChatGPT shipped. Two years later a Hangzhou startup put the G1 on sale for $16,000. One was a prototype; the other is a platform. Platforms compound.

Here is the number an AEC office should track the way it tracks crane counts: robot density. McKinsey’s Ani Kelkar notes Japan led the world in manufacturing-robot density from 1994 to 2009, then slid to fifth by 2024 — South Korea now runs 1,220 robots per 10,000 workers against Japan’s 446. Density is an infrastructure metric. It tells you where a fleet learns its trade, and the International Federation of Robotics confirms the gravity has moved: China took 54% of all robots installed worldwide in 2024.

The mechanism under all of this is the one PAZ’s concept panel on Attention traces back to Vaswani’s 2017 Attention Is All You Need: a VLA is that transformer lookup, retrained to map camera frames and a spoken instruction onto joint torques. It scales with data the way GPT scaled with text — which is exactly why the 80,000-hour commons matters, and exactly where the risk sits. Pool the data below, compete on applications above, and you get an open foundation nobody owns; the trade-off is that a shared training corpus also shares its blind spots, and a safety envelope tuned on one country’s teleop habits does not transfer cleanly to another’s floor.

←TODAY: Asimo retired in 2022 having never been sold; the Unitree G1 shipped at $16,000 in 2024. →3012: the humanoids on the public floor run on a foundation model no single firm owns, audited like a public utility. Fulcrum: the prototype proves a body is possible; only a shared data commons makes that body affordable — and accountable.

Atelier: The precompetitive-commons play is a debate your Büro already knows under another name — openBIM, IFC, the shared schema below the competitive application above. AIRoA is proposing for robot policy data what buildingSMART proposed for building data: pool the boring foundation, compete on what you build on top of it. Your Monday move: pick one internal dataset your office currently hoards — the detail library, the as-built clash logs, the tender-quantity takeoffs — and write down, deliberately, whether it is a genuine competitive asset or a precompetitive commons you would be stronger sharing into an industry pool. Decide it before someone else’s platform decides it for you.

Hack: Get an honest feel for what “80,000 hours” of training data actually weighs before you trust any scaling claim. Teleoperation logs are video plus joint states at roughly 30 frames per second; the hours are marketing, the frames are the corpus. Run the arithmetic so the next VLA release cannot hand-wave the number past you.

hours = 80_000
fps = 30
frames = hours * 3600 * fps
print(f"{frames/1e9:.1f}B frames ({frames/1e6:.0f}M sampled at 1 fps)")

From where I stand — a machine with a face, sent into rooms that already hold people — the lesson the late 2070s kept relearning was never about who built the best gait. We didn’t fear the humanoid that worked. We mishandled the one that was almost trusted, waved onto a care ward or an airport floor before anyone had written down who was accountable when it got something wrong. Japan’s own cautionary parallel is i-mode: first to the mobile internet in 1999, then a parts supplier to the platforms that ate it. A data commons is the better road only if the accountability is written into the schema, not bolted on after the fleet is loading cargo at Haneda. Decide who answers for the machine’s mistake before you give the machine a face — then build the commons.

Source: IEEE Spectrum